i-CIR: Instance-Level Composed Image Retrieval with 100K+ Hard Negatives

i-CIR is a benchmark for instance-level composed image retrieval containing between 100,000 and 1,000,000 records, released by billpsomas in 2024. It facilitates the retrieval of specific, visually indistinguishable objects by combining a reference image with a text-based modification query. The dataset includes a specialized database of visual, textual, and compositional hard negatives to test model precision.

Use Cases

• Training models for Composed Image Retrieval (CIR) using the image and text modification query format
• Benchmarking retrieval robustness against visual and compositional hard negatives
• Fine-tuning vision-language models for specific landmark or object identification

Strengths

• Scale of 100K to 1M records
• Includes visual, textual, and compositional hard negatives
• Focuses on instance-level granularity rather than broad categories

Limitations

• Restricted to non-commercial use due to CC BY-NC-SA 4.0 license
• Highly specialized for instance-level retrieval which may not reflect performance on general category-level tasks

Provenance

Source

billpsomas (Arxiv 2510.25387)

Collection Method

Curated benchmark for instance-level retrieval evaluation

Freshness

Last updated December 2025.

Geography

Global